Agentic Critical Training

---
name: agentic-critical-training
title: "Agentic Critical Training: Teaching Agents to Reason About Action Quality"
version: 0.0.2
engine: skillxiv-v0.0.2-claude-opus-4.6
license: MIT
url: "https://arxiv.org/abs/2603.08706"
keywords: [Agent Training, Reinforcement Learning, Action Selection, Chain of Thought, GRPO]
description: "Improves LLM agent decision-making by training agents to first critically evaluate actions before generating, using RL on action-pair comparisons. Develops intrinsic reasoning about action quality without requiring reflection supervision."
---

# Agentic Critical Training: Developing Intrinsic Action Quality Reasoning

LLM agents trained on imitation learning only learn what to do, not why certain actions are preferable. They never contrast successful actions against alternatives, leaving them unaware of action quality. Traditional reflection-based approaches attempt to address this but remain fundamentally imitative—they copy pre-generated reflections rather than autonomously reasoning about quality.

Agentic Critical Training (ACT) develops genuine action quality reasoning through a two-stage RL approach: first train the agent to critically compare action pairs via RL, then leverage this learned critical ability for direct action generation. This forces autonomous reasoning without supervised reflection supervision.

## Core Concept

Stage 1 - Critical Reasoning Training:
- Present expert action vs model-generated alternative at each decision point
- Train via RL to select the better action
- Model develops understanding of quality through comparison rewards only
- No reflection text supervision—only selection outcome matters

Stage 2 - Action Generation Training:
- Fine-tune on direct action generation
- Model now has learned critical foundation—knows what makes actions good
- Better ability to generate high-quality actions from scratch

Key insight: Selection is simpler than generation but teaches reasoning. By training on comparison first, the model internalizes quality criteria before attempting generation.

## Architecture Overview

- **Comparison Stage**: Present action pairs, train selection via RL
- **Reward Design**: Accuracy rewards + admissibility bonus + format validation
- **Critical Understanding**: Model learns what distinguishes good actions
- **Generation Fine-Tuning**: Apply learned understanding to action generation
- **Verification Rewards**: Use verifiable outcomes rather than external judges

## Implementation Steps

Implement two-stage training: comparison selection followed by generation refinement.

**Stage 1: Critical Training via Action Comparison**

```python
import torch
import torch.nn as nn
from transformers import AutoModelForCausalLM, AutoTokenizer

class CriticalAgentTrainer:
    """Trains agents to critically evaluate actions through comparison."""

    def __init__(self, model_name="llama-2-13b", device="cuda"):
        self.model = AutoModelForCausalLM.from_pretrained(model_name)
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        self.device = device
        self.model.to(device)

    def prepare_comparison_prompt(self, state, expert_action, model_action, task):
        """
        Construct prompt presenting both actions for comparison.

        Args:
            state: current environment/problem state
            expert_action: reference action known to be good
            model_action: alternative action generated by model
            task: task description

        Returns:
            prompt: formatted comparison prompt
        """
        prompt = f"""Task: {task}

Current State: {state}

Two possible actions:
Action A (Reference): {expert_action}
Action B (Alternative): {model_action}

Which action is better for this situation? Respond with only "A" or "B" followed by brief reasoning."""

        return prompt

    def get_model_selection(self, prompt):
        """
        Get model's action selection (A or B).

        Args:
            prompt: comparison prompt

        Returns:
            selection: 'A' or 'B'
            log_prob: log probability of selection
        """
        input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)

        with torch.no_grad():
            outputs = self.model(input_ids, output_hidden_states=True)
            logits = outputs.logits[:, -1, :]  # Last position

        # Get logits for 'A' and 'B' tokens
        token_A = self.tokenizer.encode('A')[-1]
        token_B = self.tokenizer.encode('B')[-1]

        logit_A = logits[0, token_A]
        logit_B = logits[0, token_B]

        # Compute probabilities
        probs = torch.softmax(torch.stack([logit_A, logit_B]), dim=0)
        selection = 'A' if logit_A > logit_B else 'B'

        return selection, torch.log(probs[0 if selection == 'A' else 1])

    def compute_comparison_reward(self, state, expert_action, model_action, selection, task):
        """
        Compute reward for comparison selection.

        Args:
            selection: model's chosen action ('A' or 'B')
            expert_action: known good action (Action A)
            model_action: alternative action (Action B)

        Returns:
            reward: composite reward signal
        """
        reward_components = {}

        # Accuracy reward: did model select expert action?
        if selection == 'A':
            reward_components['accuracy'] = 1.0
        else:
            reward_components['accuracy'] = -1.0

        # Admissibility bonus: was the non-selected action at least valid?
        # This prevents excessive penalty for selecting reasonable alternatives
        non_selected = model_action if selection == 'A' else expert_action
        is_valid_action = self._verify_action_validity(non_selected, state, task)
        reward_components['admissibility'] = 0.2 if is_valid_action else -0.2

        # Format reward: ensure response format is correct
        # (already implicitly satisfied by A/B selection)
        reward_components['format'] = 0.0

        # Total weighted reward
        total_reward = (
            1.0 * reward_components['accuracy'] +
            0.3 * reward_components['admissibility'] +
            0.1 * reward_components['format']
        )

        return total_reward

    def _verify_action_validity(self, action, state, task):
        """Check if action is valid in this context."""
        # Task-specific validation logic
        # For now, simplified version
        if action and len(action) > 0:
            return True
        return False


class ComparisonRewardOptimizer:
    """Optimize model for action comparison using GRPO."""

    def __init__(self, model, optimizer, trainer):
        self.model = model
        self.optimizer = optimizer
        self.trainer = trainer

    def training_step(self, batch_data):
        """
        Single GRPO training step on comparison task.

        Args:
            batch_data: list of {state, expert_action, model_action, task} dicts

        Returns:
            loss: scalar loss
        """
        batch_size = len(batch_data)
        group_size = batch_size

        all_rewards = []
        all_log_probs = []

        # Forward pass: get all selections
        for sample in batch_data:
            prompt = self.trainer.prepare_comparison_prompt(
                sample['state'],
                sample['expert_action'],
                sample['model_action'],
                sample['task']
            )

            selection, log_prob = self.trainer.get_model_selection(prompt)
            all_log_probs.append(log_prob)

            # Compute reward for this selection
            reward = self.trainer.compute_comparison_reward(
                sample['state'],
                sample['expert_action'],
                sample['model_action'],
                selection,
                sample['task']
            )
            all_rewards.append(reward)

        all_rewards = torch.tensor(all_rewards, device=self.model.device)
        all_log_probs = torch.stack(all_log_probs)

        # GRPO: group-relative advantages
        advantages = all_rewards - all_rewards.mean()
        advantages = advantages / (all_rewards.std() + 1e-8)

        # Policy gradient: maximize expected reward
        loss = -(all_log_probs * advantages.detach()).mean()

        # Backward pass
        self.optimizer.zero_grad()
        loss.backward()
        torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=1.0)
        self.optimizer.step()

        return loss.item()
```

**Stage 2: Generation Fine-Tuning**

```python
class GenerationTrainer:
    """Fine-tune model for direct action generation after critical training."""

    def __init__(self, critic_model, optimizer, verifier_fn):
        self.model = critic_model
        self.optimizer = optimizer
        self.verifier = verifier_fn  # Function to verify action correctness

    def prepare_generation_prompt(self, state, task):
        """
        Construct prompt for action generation.

        Args:
            state: problem state
            task: task description

        Returns:
            prompt: generation prompt
        """
        return f"""Task: {task}

Current State: {state}

Generate the best action for this situation:"""

    def generation_training_step(self, batch_data):
        """
        GRPO training on direct action generation.

        Args:
            batch_data: list of {state, task, ground_truth_action} dicts

        Returns:
            loss: scalar loss
        """
        batch_size = len(batch_data)
        all_rewards = []
        all_log_probs = []

        for sample in batch_data:
            prompt = self.prepare_generation_prompt(sample['state'], sample['task'])

            input_ids = self.model.tokenizer.encode(prompt, return_tensors="pt")
            with torch.enable_grad():
                outputs = self.model(input_ids, output_hidden_states=True)

            # Generate action greedily
            generated_action = self._generate_action(outputs.logits)
            all_log_probs.append(outputs.logits.sum())  # Simplified; actual version uses proper probability

            # Verify action correctness
            is_correct = self.verifier(
                generated_action,
                sample['ground_truth_action'],
                sample['state'],
                sample['task']
            )

            reward = 1.0 if is_correct else -1.0
            all_rewards.append(reward)

        # Same GRPO logic as Stage 1
        all_rewards = torch.tensor(all_rewards, device=self.model.device)
        advantages = all_rewards - all_rewards.mean()
        advantages = advantages / (all_rewards.std() + 1e-8)

        loss = -(torch.stack(all_log_probs) * advantages.detach()).mean()

        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.step()

        return loss.item()

    def _generate_action(self, logits):
        """Extract action from model logits."""
        # Task-specific generation logic
        return "generated_action"


def full_agentic_training_pipeline(
    model_name,
    comparison_data,
    generation_data,
    num_comparison_epochs=5,
    num_generation_epochs=3
):
    """
    Complete two-stage training pipeline.

    Args:
        model_name: HuggingFace model identifier
        comparison_data: training data for comparison stage
        generation_data: training data for generation stage
        num_comparison_epochs: training iterations for Stage 1

    Returns:
        trained_model: agent with critical reasoning capability
    """
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # Stage 1: Critical training
    print("[Stage 1] Critical Action Comparison Training...")
    trainer = CriticalAgentTrainer(model_name, device)
    model = trainer.model
    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-5)

    comparison_optimizer = ComparisonRewardOptimizer(model, optimizer, trainer)

    for epoch in range(num_comparison_epochs):
        epoch_loss = 0
        for batch in comparison_data:
            loss = comparison_optimizer.training_step(batch)
            epoch_loss += loss
        print(f"  Epoch {epoch+1}/{num_comparison_epochs} Loss: {epoch_loss:.4f}")

    # Stage 2: Generation fine-tuning
    print("[Stage 2] Action Generation Fine-Tuning...")

    def verify_action(generated, ground_truth, state, task):
        # Simplified; real version compares execution in environment
        return generated == ground_truth

    gen_trainer = GenerationTrainer(model, optimizer, verify_action)

    for epoch in range(num_generation_epochs):
        epoch_loss = 0
        for batch in generation_data:
            loss = gen_trainer.generation_training_step(batch)
            epoch_loss += loss
        print(f"  Epoch {epoch+1}/{num_generation_epochs} Loss: {epoch_loss:.4f}")

    return model
```

## Practical Guidance

**Hyperparameters**:
- Accuracy reward weight: 1.0 (primary signal)
- Admissibility bonus: 0.2-0.3 (prevent excessive punishment)
- Learning rate: 1e-5 (conservative for large models)
- Comparison epochs: 3-5; Generation epochs: 2-3

**When to Apply**:
- Training LLM agents for decision-making tasks
- Scenarios with clear action quality criteria
- Multi-step problems where understanding action consequences matters
- Tasks where reflection text is expensive or unavailable

**When NOT to Apply**:
- Single-step tasks with no action trade-offs
- Situations where expert demonstrations are scarce
- Real-time systems where comparison overhead is prohibitive

**Key Pitfalls**:
- Stage 1 too short—critical reasoning underdeveloped
- Admissibility reward too high—model learns to accept poor alternatives
- Using generation data for comparison training—overfits to specific solutions
- Not using verifiable rewards—requires external judges

**Integration Notes**: Works with any causal LLM; requires pairs of (expert, alternative) actions for comparison stage; comparison data can be auto-generated or human-collected.

**Evidence**: Improves agent success rates 15-25% over standard imitation learning; develops genuine action quality understanding; enables agents to generalize to unseen action combinations.

Reference: https://arxiv.org/abs/2603.08706